By Jason An, Ph.D., CEO, Medicus
Last updated: 2026-10

The Tumor Proportion Score (TPS) is the percentage of viable tumor cells with PD-L1 membrane staining, while the Combined Positive Score (CPS) counts PD-L1-stained tumor cells, lymphocytes and macrophages, divides by all viable tumor cells and multiplies by 100. In gastric and gastroesophageal junction (GEJ) adenocarcinoma, CPS is the score used to select patients for anti-PD-1 therapy, and its cutoff (≥1 or ≥5) depends on the drug, the assay and the region.

This guide is educational and is not clinical advice. Treatment decisions must follow the current local drug label, the instructions for use of the validated assay, and the judgment of the treating physician and pathologist.

TPS and CPS at a glance

Both scores are read on the same type of slide — an immunohistochemistry (IHC) section stained with a PD-L1 antibody — but they ask different questions.

TPS asks: what fraction of the tumor cells express PD-L1? Agilent's PD-L1 IHC 22C3 pharmDx interpretation manual for non-small cell lung cancer (NSCLC) defines TPS as the percentage of viable tumor cells showing partial or complete membrane staining at any intensity. Only tumor-cell membrane staining counts; cytoplasmic staining and immune cells are excluded, and the specimen must contain at least 100 viable tumor cells [2].

CPS asks: how much PD-L1 is present in the tumor and its immune infiltrate, relative to the tumor size? The 22C3 gastric/GEJ interpretation manual counts PD-L1-staining tumor cells, lymphocytes and macrophages, divides by the total number of viable tumor cells, and multiplies by 100; a specimen is considered PD-L1 positive at CPS ≥1 [1]. Because immune cells sit in the numerator, the arithmetic can exceed 100, so Agilent's 22C3 scoring guides cap the reported score at CPS 100 [16].

Why gastric cancer moved to CPS

CPS was developed for gastric cancer because TPS performed poorly there. In an early pembrolizumab study (KEYNOTE-012), a tumor-cell-only score identified just 2 of 11 gastric responders, prompting a score that also captures PD-L1 on immune cells [15]. Gastric tumors often express PD-L1 more on infiltrating immune cells than on tumor cells, which a tumor-only count would miss.

What counts — and what doesn't — in gastric CPS

The numerator rules are where most scoring errors begin. A 2025 best-practice review from Korean gastrointestinal pathologists summarizes them [11]:

• Included: tumor cells with membrane staining (partial or complete, any intensity); lymphocytes and macrophages with membrane and/or cytoplasmic staining.
• Excluded: tumor cells with cytoplasmic-only staining; PD-L1-positive benign glands or intestinal metaplasia; fibroblasts, smooth muscle and endothelial cells; neutrophils, eosinophils and plasma cells; necrotic or mucinous debris; ulcer and gastritis areas.
• Biopsy heterogeneity: small endoscopic biopsies may not represent the tumor; the review cites sampling at least four to five biopsy fragments.
• Assay differences: in the authors' prior series, the 28-8 assay showed non-specific staining (mostly muscle or glands) more often than 22C3, which must not be counted [11].

Gastric cancer cutoffs: drug, assay and region

Cutoffs are not universal. They are tied to a specific drug, a specific trial and a specific regulatory jurisdiction — and they have changed over time.

Key context behind the table:

• CheckMate-649 randomized 1,581 patients; its dual primary endpoints were overall and progression-free survival in patients with CPS ≥5, scored with the 28-8 assay [7]. The original 2021 US approval of nivolumab was irrespective of PD-L1; the current US label limits use to tumors expressing PD-L1 (≥1) [4][7].
• KEYNOTE-859 randomized 1,579 patients with HER2-negative disease; median overall survival was 12.9 vs 11.5 months (HR 0.78), with prespecified analyses at CPS ≥1 and CPS ≥10 [9]. The current US Keytruda label restricts this indication to CPS ≥1 [3].
• At a September 2024 FDA Oncologic Drugs Advisory Committee meeting, the agency's analysis described a "consistent lack of benefit" in the PD-L1 low (CPS <1) group across anti-PD-1 gastric trials [10].
• For HER2-positive disease, the FDA amended the pembrolizumab indication on November 7, 2023 to CPS ≥1 and approved 22C3 pharmDx as its companion diagnostic [5]. On July 14, 2026, Agilent announced FDA approval of PD-L1 IHC 28-8 pharmDx as a diagnostic aid at CPS ≥1 for nivolumab in gastric, GEJ and esophageal cancers [6]; laboratories should confirm that the current label and the FDA companion diagnostic list reflect this.

Because assays, cutoffs and reimbursement rules differ, a single slide can be "positive" under one framework and "negative" under another. Reports should state the antibody clone, the scoring method and the cutoff applied.

The consistency problem

CPS is harder to reproduce than it looks. Identifying every PD-L1-positive lymphocyte and macrophage, excluding plasma cells and neutrophils, and counting all viable tumor cells across a whole slide is a large manual estimation task.

• In a study published in Modern Pathology (2023), 12 pathologists scored 100 gastroesophageal adenocarcinoma biopsies under standardized conditions; interobserver agreement for CPS was low to moderate and did not improve after training [13].
• In a companion real-world study, 14 pathologists at 13 institutions scored 108 gastric/GEJ cases; overall agreement at the CPS <1 vs ≥1 cut point was about 30%, with better agreement at higher cutoffs [13].
• At a Korean consensus meeting, 11 expert GI pathologists scored 20 slides with CPS near 5 (range 3–7); only a minority of cases showed good agreement, and one case reached no consensus [12].
• The 2025 best-practice review notes that inter-reader variability is greatest near the CPS 5 cutoff, driven by faint staining, membrane-versus-cytoplasm calls and crush or cautery artifacts [11].

Variability is largest exactly where it matters most: close to the decision threshold.

How AI improves CPS consistency

AI does not change the definition of CPS. It changes how the counting is done — exhaustively, cell by cell, and the same way every time.

A typical AI CPS pipeline works in stages:

1. Tissue and tumor-region detection on the whole-slide image (WSI), excluding background, necrosis and non-tumor mucosa.
2. Cell detection and classification into PD-L1-positive and -negative tumor cells, and PD-L1-positive lymphocytes/macrophages.
3. CPS calculation from whole-slide counts, with region-level heatmaps so the pathologist can see where the score comes from.

Published evidence in gastric cancer is still early but encouraging. Han et al. (2025) built a multi-model pipeline on 308 gastric cancer WSIs stained with 22C3; AI-derived CPS agreed with pathologist consensus with Cohen's kappa of 0.782 in an internal test set and 0.737 in an external cohort [14]. The authors are explicit about limits — two institutions, one scanner type, one antibody, misses on faint membranes, and false positives from mucin and debris — and position the tool as an adjunct to, not a replacement for, the pathologist [14]. The 2025 best-practice review likewise notes that AI-assisted PD-L1 assessment has been validated mainly in lung cancer and remains less explored in gastric cancer [11].

What AI can realistically offer today:

• Reproducibility: the same slide yields the same count.
• Exhaustive counting: whole-slide counts instead of field-by-field estimates.
• Transparency near cutoffs: heatmaps and cell overlays focus pathologist review on borderline regions.
• Audit trail: stored counts support QA, proficiency testing and research.

Any AI used in a clinical workflow must be validated for the specific antibody, staining platform, scanner and indication, and used within its regulatory status in each country.

How Medicus approaches this

Medicus, Inc. develops cell-based AI for pathology. PD-L1 interpretation is subjective and can vary by tissue region, which makes quantitative, consistent assessment difficult. MeDICus® is our biomarker AI designed to support quantitative, consistent PD-L1 assessment through AI quantification; its regulatory status is not described here as an approval or clearance, and it is presented for research use. Our models draw on gastric and breast cancer clinical datasets built and validated with tertiary hospitals in Korea.

Reliable AI scoring starts with organized image data. Many teams first consolidate whole-slide images from multiple scanners with MeDIAuto®, our pathology data infrastructure. To discuss a PD-L1 research collaboration or pilot, contact our team.

FAQ

What is the difference between CPS and TPS?

TPS counts only PD-L1-positive tumor cells as a percentage of viable tumor cells. CPS adds PD-L1-positive lymphocytes and macrophages to the numerator, divides by viable tumor cells, multiplies by 100 and caps at 100 [1][2][16].

Why is CPS used in gastric cancer instead of TPS?

In early pembrolizumab data, a tumor-cell-only score missed most gastric responders (2 of 11 identified), so CPS was developed to capture immune-cell PD-L1 as well [15]. Current US labels for pembrolizumab and nivolumab in gastric cancer use CPS-based selection [3][4].

Is the gastric cancer cutoff CPS ≥1 or CPS ≥5?

It depends on drug and region. In the US, current pembrolizumab and nivolumab labels use ≥1 [3][4]. The EU nivolumab indication uses CPS ≥5 [8], and Korean nivolumab reimbursement was limited to CPS ≥5 as of 2024 [12]. Always check the current local label.

Are the 22C3 and 28-8 assays interchangeable?

Not automatically. Each was developed with a specific drug (22C3 with pembrolizumab, 28-8 with nivolumab), and 28-8 shows non-specific staining more often [11]. Use the assay specified by the label or a properly validated alternative.

How much do pathologists disagree on CPS?

Substantially in some studies: about 30% overall agreement at the CPS 1 cut point among 14 pathologists, and low-to-moderate agreement among 12 pathologists that did not improve with training [13]. Variability is greatest near cutoffs [11].

Can AI replace the pathologist for PD-L1 scoring?

No. Current studies position AI as an adjunct that improves reproducibility and focuses review on borderline regions; the pathologist remains responsible for the final interpretation [11][14].

References

1. Agilent Technologies. PD-L1 IHC 22C3 pharmDx Interpretation Manual – Gastric or Gastroesophageal Junction Adenocarcinoma (US). agilent.com (PDF)
2. Agilent Technologies. PD-L1 IHC 22C3 pharmDx Interpretation Manual – NSCLC (KEYNOTE-042). agilent.com (PDF)
3. KEYTRUDA (pembrolizumab) Prescribing Information, revised 07/2026. DailyMed
4. OPDIVO (nivolumab) Prescribing Information, revised 08/2026. DailyMed
5. US FDA. FDA amends pembrolizumab's gastric cancer indication (Nov 7, 2023). fda.gov
6. Agilent Technologies. Agilent Receives FDA Approval for PD-L1 IHC 28-8 pharmDx in ESCC, Gastric, GEJ and Esophageal Adenocarcinoma (Jul 14, 2026). investor.agilent.com
7. Janjigian YY, Shitara K, Moehler M, et al. First-line nivolumab plus chemotherapy versus chemotherapy alone for advanced gastric, gastro-oesophageal junction, and oesophageal adenocarcinoma (CheckMate 649). Lancet. 2021;398:27-40. Summary: The ASCO Post
8. European Medicines Agency. CHMP post-authorisation summary of positive opinion for Opdivo (II/96). ema.europa.eu (PDF)
9. Rha SY, et al. Pembrolizumab plus chemotherapy versus placebo plus chemotherapy for HER2-negative advanced gastric cancer (KEYNOTE-859). Lancet Oncol. 2023. Summary: The ASCO Post
10. US FDA. Oncologic Drugs Advisory Committee, Sept 26, 2024 – FDA presentation on PD-L1 expression and anti-PD-1 therapy in HER2-negative gastric/GEJ adenocarcinoma. fda.gov
11. Ahn S, Hwang I, Kim Y, et al. Best Practice PD-L1 Staining and Interpretation in Gastric Cancer Using PD-L1 IHC PharmDx 22C3 and 28-8 Assays. Biomedicines. 2025;13(11):2824. PMC12649831
12. Ahn S, Kwak Y, Kwon GY, et al. Interpretation of PD-L1 expression in gastric cancer: summary of a consensus meeting of Korean gastrointestinal pathologists. J Pathol Transl Med. 2024;58(3):103-116. KoreaMed Synapse
13. Yale School of Medicine. Research from Yale Pathology highlights the challenges of interpreting PD-L1 expression with CPS method (summarizing Robert ME et al. and Rimm DL et al., Mod Pathol 2023). medicine.yale.edu
14. Han T, Zhuo M, Song Z, et al. Automated interpretation of PD-L1 CPS based on multi-AI models integration strategy in gastric cancer. Front Immunol. 2025;16:1614099. frontiersin.org
15. Paxton A. Scoring gastric, GEJ cancers for PD-L1 expression. CAP TODAY. Feb 2018. captodayonline.com
16. Agilent Technologies. PD-L1 IHC 22C3 pharmDx Quick Reference – Esophageal Cancer (CPS calculation; maximum score CPS 100). agilent.com (PDF)